Open K-King6 opened 2 weeks ago
Thank you for pointing out this issue, which brought to our attention that the published version is the variant of "what -> when -> where". If you intend to use it as a baseline, we appreciate it if you could revert it to the 'when -> what -> where' version proposed in the paper. Thank you once again and best wishes.
cur_t_rnn, hc_t = self.capturer_t(rnn_input_his_concat, rnn_input_cur_concat, his_mask, cur_mask, mask_batch[1:]) if self.cat_contained: cur_c_rnn, hc_c = self.capturer_c(rnn_input_his_concat, rnn_input_cur_concat, his_mask, cur_mask, mask_batch[1:], hc_t) cur_l_rnn, hc_l = self.capturer_l(rnn_input_his_concat, rnn_input_cur_concat, his_mask, cur_mask, mask_batch[1:], hc_c)
You first calculate hc_t and use it to calculate hc_c,but you then first calculate c_trans and use it to calculate t_pred, it seems not consistent and may make your result worse.![Snipaste_2024-06-20_20-27-54](https://github.com/urbanmobility/CSLSL/assets/87562118/439a1815-e5bc-4c9c-a0b1-816afd2699a6)